optimizing-parser-performance

name	optimizing-parser-performance
description	Audits parser and tokenizer performance.
metadata	{"internal":true}

Optimizing Parser Performance

Profile release-like parsing against representative valid, malformed, small, large, ASCII, and Unicode corpora. Preserve grammar behavior, diagnostics, source locations, and resource limits.

1. Read the shared parser data-oriented design audit.
2. Scan for copied token text, allocation per node/token, pointer-rich ASTs, scattered passes, generic Unicode work in ASCII-heavy scans, and JS/native/Wasm crossings per item.
3. Change one causal storage or scan-loop cost at a time; define ownership, offsets, and overflow behavior before compacting data.
4. Re-measure CPU, allocations, peak memory, startup, and throughput; test malformed and deeply nested input alongside the common corpus.

Use optimizing-compiler-performance only for a profile-proven codegen issue.

SIMD scan loops

A lexer scanning identifiers, strings, template bodies, or whitespace is the case that pays off for hand SIMD — the optimizer will not autovectorize a data-dependent find-first-of-a- byte-set scan, so reach for explicit SIMD or a memchr-family primitive. Portable SIMD is runtime CPU dispatch; do not ship -C target-cpu=native or GOAMD64=v2|v3|v4 in a build distributed to CPUs you do not control. Pin only to a floor a controlled target guarantees, and record why, enforced by scripts/fleet/check/build-microarch-is-portable.mts.

• The kernel is compare-and-reduce. Load a 16/32-byte chunk, run the class compares, OR the class masks, extract to a scalar bitmask, find the first boundary with a count-trailing-zeros — or NOT then count for the first NON-member — then a scalar tail handles the sub-stride remainder.
• Byte-identical or it does not ship. A SIMD scan must match its scalar reference exactly: ship a SIMD-vs-scalar differential test plus an exhaustive delimiter-at-every- offset-across-the-stride test, and validate end to end.
• The SIMD candidate byte set must match its OWN scalar path exactly. It may be a SUPERSET of a sister port's set when this port carries extra semantics — e.g. a Go readString scanning 0xE2 for U+2028/U+2029 handling that the Rust and C++ ports omit. Correctness beats cross-port delimiter symmetry.
• Make it real and wire it in. An unrolled scalar loop with a per-byte call is NOT SIMD; a SIMD helper with no non-test caller is dead code. A shipped SIMD function must be a real vector kernel AND live on the hot path — prefer whatever the sister ports do, in lock-step.
• Measured payoff shape. The acorn-lang Go work landed 1.5-2.4x end-to-end lexer throughput on string/template-dense JS, ~16x on isolated micro-scans, allocation-neutral. Language mechanics: Rust, Go, C++.